The Agent Skills Directory

[COMMAND_EXECUTION]: The skill uses Python's subprocess module to interact with the system environment. This includes executing the platform's command-line interface (claude) to run evaluations and improvement loops, and using lsof to manage local network ports for the visualization tool.
[EXTERNAL_DOWNLOADS]: The visualization components (assets/eval_review.html, eval-viewer/viewer.html, and scripts/generate_report.py) reference external assets from well-known and trusted services, including Google Fonts and the SheetJS library via cdn.sheetjs.com. These are used for styling and processing spreadsheet data in reports.
[INDIRECT_PROMPT_INJECTION]: The skill acts as a meta-tool that processes other, potentially untrusted, skill instructions and descriptions provided by the user.
Ingestion points: Reads skill definitions and evaluation results from the workspace via scripts/utils.py and scripts/improve_description.py.
Boundary markers: Prompts utilize XML-style tags (e.g., <skill_content>, <new_description>) to isolate untrusted skill data from the optimization instructions.
Capability inventory: The skill can execute CLI commands via claude -p, write file system changes (e.g., scripts/package_skill.py), and host a local web server for reviewing results.
Sanitization: Uses YAML block scalars when generating temporary configuration files to prevent formatting breakages, and employs delimited prompting to guide the assistant's behavior.

create-a-skill